Contextual Bandits

Contextual bandits are bandit problems where each decision arrives with observable context, and the algorithm learns which actions work best for which contexts from partial feedback.

Key points

• Slivkins describes contextual bandits as a middle ground between IID and adversarial bandits: rewards can change across rounds, but the change is explained by observed contexts ^[src-019].
• The model is useful when actions should be personalised or conditioned on features, such as recommending content, choosing offers, or adapting an interface to a user/session context ^[src-019].
• The learning problem is harder than Stochastic Bandits because the algorithm must learn a policy from contexts to actions, not just estimate one reward distribution per arm ^[src-019].
• Contextual bandits sit near reinforcement learning conceptually but keep the feedback loop shorter: each round is a context, action, and observed reward rather than a long-horizon Markov decision process ^[src-019].
• Lu, Pál, and Pál make the model concrete through sponsored search: a query arrives as context, an ad is chosen as the action, and click feedback is the observed payoff ^[src-020].
• Their Lipschitz version adds a metric structure over contexts and actions, allowing reward estimates to generalise across similar queries and similar ads ^[src-020].
• Yildirim frames contextual bandits as dynamic Treatment Personalisation: traffic allocation changes over time and by user context, not just globally as in a context-free bandit ^[src-021].
• A contextual bandit is often the practical middle ground between static A/B testing and full multi-step reinforcement learning: useful when actions do not materially change future system state ^[src-021].
• In production experimentation, contextual bandits also need Offline Policy Evaluation so teams can test candidate policies against logged data before live deployment ^[src-021].
• Hightouch frames contextual bandits as the personalisation extension of standard bandits: instead of finding the single best option for everyone, they use customer context such as purchase history and demographic information to choose for the individual ^[src-023].
• Hightouch's MAB article repeats the limitation from the other side: standard multi-armed bandits find winning strategies faster but still optimise for the average customer, so contextual bandits are needed for Sarah-versus-Marcus style individual differences ^[src-025].
• Hightouch's contextual-bandit article makes the operational pipeline explicit: customer data becomes a Customer Feature Matrix, possible actions are combined across dimensions, a model predicts expected reward, and the bandit selects the highest-reward action for that individual ^[src-026].
• The article emphasises dual learning: audience-level patterns help similar customers and cold-start cases, while individual-level updates preserve exceptions such as a high-LTV customer who ignores discounts ^[src-026].
• At enterprise scale, contextual-bandit systems must process hundreds of customer features, thousands of possible actions, coordinate multiple campaigns, avoid overwhelming customers, and connect with existing ML models ^[src-026].
• Braze describes contextual bandits as an added intelligence layer over standard bandits, using signals such as location, device, and past engagement to choose what performs best for a customer in a specific moment ^[src-027].
• In Braze's AI Decisioning Studio framing, contextual bandits work alongside Intelligent Selection and predictive scoring rather than replacing the MAB allocation layer ^[src-027].
• Statsig positions contextual multi-armed bandits as an acceleration tool for shallow A/B decisions where the rule is binary, such as picking the winner and killing the loser based on conversion rate ^[src-031].
• Statsig also cautions that contextual bandits are less appropriate for careful calibration studies where inference quality matters more than traffic shifting ^[src-031].
• Braze describes contextual bandits as having two parts: an oracle that predicts expected reward and an exploration strategy that decides when to test uncertain actions ^[src-152].
• Braze says its contextual bandits use customer features, action features, and environment signals, and its whitepaper names elastic nets and gradient-boosted decision trees as model families for navigating the bias-variance tradeoff in marketing decisioning ^[src-152].
• The Braze source also surfaces a practical governance issue: customer context can create unfair or illegal outcomes when features act as proxies for protected characteristics, so decisioning systems need explicit use-case judgement and safeguards ^[src-152].

Related entities

• Aleksandrs Slivkins
• Braze

Related concepts

• Multi Armed Bandits
• Exploration Exploitation Tradeoff
• Stochastic Bandits
• Adversarial Bandits
• Thompson Sampling
• Lipschitz Contextual Bandits
• Query Ad Clustering
• Sponsored Search Ad Ranking
• Treatment Personalisation
• Epsilon Greedy
• Upper Confidence Bound
• Offline Policy Evaluation
• AI Decisioning
• Personalisation Gap
• Marketing Bandit Optimisation
• Customer Feature Matrix
• Intelligent Selection
• Ab Test Acceleration
• Sequential Testing
• Community Of Bandits

Source references

• ^[src-019] Aleksandrs Slivkins — "Introduction to Multi-Armed Bandits" (2019-04-15; revised 2024-04-03)
• ^[src-020] Tyler Lu, David Pál, Martin Pál — "Contextual Multi-Armed Bandits" (AISTATS 2010)
• ^[src-021] Ugur Yildirim — "An Overview of Contextual Bandits" (2024-02-02)
• ^[src-023] Hightouch — "Under the hood of AI Decisioning, part one: Overcoming the personalization gap"
• ^[src-025] Hightouch — "Under the hood of AI Decisioning, part three: Multi-armed bandits"
• ^[src-026] Hightouch — "Under the hood of AI Decisioning, part four: Contextual bandits"
• ^[src-027] Team Braze — "What is a multi-armed bandit? Smarter experimentation for real-time marketing"
• ^[src-031] Yuzheng Sun — "Speeding up A/B tests with discipline"
• ^[src-152] George Khachatryan, Nathaniel Rounds, Victor Kostyuk / Braze – "A community of bandits" (source page: "From multivariate testing to AI decisioning", 2025-09-30)